Research On Dense Visual Simultaneous Localization And Mapping In Dynamic Environments

Visual simultaneous localization and mapping(VSLAM)is the key technology of robot systems such as unmanned aerial vehicle(UAV),augmented reality(AR),unmanned driving and human-computer interaction,which has a idealistic hypothesis of static environment.It is inevitable that moving objects would be probably appeared in the actual application scenarios,especially some objects with obvious movement that will against the static environment assumption,which makes the available VSLAM systems difficult to meet the application requirements of the actual scenarios.Compared with feature-based VSLAM method,dense VSLAM method has better robustness and stability in motion blur or featureless regions.However,dense VSLAM method is more susceptible to dynamic environments.In dynamic scenes,the available dense VSLAM method faces two challenges.On one hand,motion segmentation,as the key algorithm to solve the dynamic scenes,which the accuracy and robustness in the complex dynamic scene is still a challenging problem.The available geometric information based segmentation method is difficult to adapt to the highly dynamic scenes,while the instance semantic segmentation based method is easy to cause false positive over-segmentation error,and resulting degradation of performance of the pose estimation based on the direct method.On the other hand,for the dense visual odometry,if the dynamic object participates in the matching process of dense visual odometry,it will greatly deteriorate the performance of camera pose estimation.With regard to the dense VSLAM system,the dynamic object will have influence on the accuracy of loop detection,thus leading to wrong loop and reducing its success rate.The dynamic objects will also have great damage to the pose graph model,and resulting in wrong pose edge association,which degrade the performance of the pose graph optimization.To solve the above problems,this thesis deeply studies the problem of VSLAM in the dynamic environments,and propose the mathematical model of dense VSLAM in dynamic environment based on motion segmentation,and the corresponding dense VSLAM methods.Specifically,the main contributions include:(1)The mathematical model of dense VSLAM for dynamic environments is established.Based on the mathematical model of dense VSLAM in static environments,the specific influence of moving objects on each module of dense VSLAM is analyzed.The mathematical model of dense VSLAM in dynamic environments is described with motion segmentation results,which provides theoretical basis for dense VSLAM to deal with dynamic scenes.(2)A dense visual odometry method in dynamic environments is proposed.The scene clustering is performed based on the combination of the image depth and gray information.According to the matching residual between image frames,a multiple frame accumulated residual model is constructed.Then the nonparametric statistical model of residual distribution is built,and the motion segmentation model based on dynamic threshold,as well as the cluster-based mixed weight model,is constructed by using the statistical model.Finally,the motion segmentation results and the clusteringbased weight model are added to the energy-based optimization function for pose estimation,and achieve a more robust camera motion estimation than the avilable dense methods in dynamic environments.(3)A robust dense VSLAM method fused semantic segmentation in dynamic environments is proposed,which combines residual-based segmentation and instance segmentation to achieve more accurate and robust scene motion segmentation than geometric-based methods.With the help of motion segmentation,a dense visual odometry with frame-keyframe matching in dynamic environments is constructed by introducing the keyframe mechanism,which effectively reduces the local cumulative error.Combined with motion segmentation results,the loop closure and pose graph optimization model in dynamic environments is built,which further reduces the cumulative error of trajectory estimation and the influence of moving objects on camera pose estimation.The proposed dense VSLAM method achieves better performance than the other dense VSLAM methods in dynamic environments.